Case studies
Three real optimization runs against public repos — one verified, one blocked before it could ship, one discounted by our own verification discipline before it ever became a claim. We publish all three the same way, because a guarantee only means something if the machinery behind it is willing to throw out a number that doesn't survive contact with reality.
The estimate said +45.6%. Ten interleaved runs on real GitHub-hosted CI measured +14.1%, statistically significant. The change was real; the estimate ran 3× hot — and the measurement caught it.
Read the case study →A genuine, load-bearing opportunity confirmed across 5 independent real runs — but the verification path is structurally blocked (no fork access to the triggering branch or the required secrets). Held back rather than shipped on an estimate alone.
Read the case study →A clean, webpack-tier estimate that looked identical to a real win on paper — until we checked it against the repo's actual structure. Half was unshippable to any fork, the other half was already sharded 108 ways. The discipline killed it before it became a promise.
Read the case study →